Abstract
Dating ancient artworks is a fundamental challenge in art history, traditionally relying on subjective visual analysis. This research introduces a quantitative framework for dating mural paintings from the Mogao Grottoes by extracting "artistic style" through a combination of handcrafted features and deep learning. Using a custom dataset of 3,860 paintings across six dynasties, the proposed hybrid model achieves high classification accuracy and offers a reproducible method for resolving archaeological controversies.
Introduction
The Mogao Grottoes in Dunhuang, China, house a vast collection of Buddhist art spanning a millennium. While many caves are well-documented, others lack historical records, leading to intense academic debate regarding their age. Traditional "connoisseurship" relies on stylistic cues—such as the transition from slim figures in the Sui Dynasty to more robust forms in the Tang Dynasty. However, this human-centric approach is inherently subjective. This study aims to transform qualitative stylistic analysis into a quantitative science using a hybrid computer vision framework.
Methodology
The researchers treat the dating task as a drawing-style classification problem. To capture the multifaceted nature of "style," they developed a multi-pronged feature extraction approach:
1. Handcrafted Descriptors
- Shape: Scale-Invariant Feature Transform (SIFT) was used to capture brushstroke geometry, encoded via an Improved Fisher Vector (IFV) for a global representation of line work.
- Color: The Regional Color Co-occurrence Descriptor (RCC) was employed to analyze color collocation—how colors are positioned relative to one another—which proved more indicative of era than simple color distribution.
2. Deeply Learnt Visual Codes (DunNet)
A custom deep convolutional neural network, DunNet, was designed to identify implicit stylistic patterns. High-level "visual codes" were extracted from the network's second fully connected layer to represent complex stylistic nuances.
3. Multi-Scale Voting
To account for varying mural sizes, the system crops 100 sub-images at multiple scales from each mural. A majority-vote strategy across these samples determines the final chronological label.
Datasets and Evaluation
The study utilized two key datasets: the DunHuang-E6 benchmark, containing 3,860 paintings from 194 grottoes across six dynasties, and the DunHuang-P7 prediction set, consisting of seven murals from grottoes with historically debated dates.
Results and Discussion
The combined hybrid model achieved a 71.64% accuracy on the 6-class classification task. Performance improved significantly in binary comparisons between neighboring dynasties, often exceeding 88%. The most challenging distinction was between the Middle and Late Tang periods, reflecting the high stylistic similarity of those eras.
Beyond metrics, the system provided objective evidence for real-world archaeological disputes. In Grotto #206, the model confirmed a Sui Dynasty style despite its location among Tang-era caves. Furthermore, the analysis of a "Peacock King" painting suggested a Middle Tang origin rather than the previously assumed Wu Dai period, prompting new academic inquiry.
Conclusion
By quantifying artistic style, this research bridges the gap between traditional art history and modern computer vision. The success of the hybrid approach—combining deep learning with specific color and shape descriptors—demonstrates that "style" is a complex construct of both low-level textures and high-level semantic patterns. This framework provides a powerful, objective tool for the preservation and study of global cultural heritage.
Summary updated